xref: /dpdk/drivers/net/mlx5/mlx5.c (revision 68a03efeed657e6e05f281479b33b51102797e15)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox Technologies, Ltd
4  */
5 
6 #include <stddef.h>
7 #include <unistd.h>
8 #include <string.h>
9 #include <stdint.h>
10 #include <stdlib.h>
11 #include <errno.h>
12 
13 #include <rte_malloc.h>
14 #include <ethdev_driver.h>
15 #include <ethdev_pci.h>
16 #include <rte_pci.h>
17 #include <rte_bus_pci.h>
18 #include <rte_common.h>
19 #include <rte_kvargs.h>
20 #include <rte_rwlock.h>
21 #include <rte_spinlock.h>
22 #include <rte_string_fns.h>
23 #include <rte_alarm.h>
24 #include <rte_cycles.h>
25 
26 #include <mlx5_glue.h>
27 #include <mlx5_devx_cmds.h>
28 #include <mlx5_common.h>
29 #include <mlx5_common_os.h>
30 #include <mlx5_common_mp.h>
31 #include <mlx5_common_pci.h>
32 #include <mlx5_malloc.h>
33 
34 #include "mlx5_defs.h"
35 #include "mlx5.h"
36 #include "mlx5_utils.h"
37 #include "mlx5_rxtx.h"
38 #include "mlx5_autoconf.h"
39 #include "mlx5_mr.h"
40 #include "mlx5_flow.h"
41 #include "mlx5_flow_os.h"
42 #include "rte_pmd_mlx5.h"
43 
44 /* Device parameter to enable RX completion queue compression. */
45 #define MLX5_RXQ_CQE_COMP_EN "rxq_cqe_comp_en"
46 
47 /* Device parameter to enable padding Rx packet to cacheline size. */
48 #define MLX5_RXQ_PKT_PAD_EN "rxq_pkt_pad_en"
49 
50 /* Device parameter to enable Multi-Packet Rx queue. */
51 #define MLX5_RX_MPRQ_EN "mprq_en"
52 
53 /* Device parameter to configure log 2 of the number of strides for MPRQ. */
54 #define MLX5_RX_MPRQ_LOG_STRIDE_NUM "mprq_log_stride_num"
55 
56 /* Device parameter to configure log 2 of the stride size for MPRQ. */
57 #define MLX5_RX_MPRQ_LOG_STRIDE_SIZE "mprq_log_stride_size"
58 
59 /* Device parameter to limit the size of memcpy'd packet for MPRQ. */
60 #define MLX5_RX_MPRQ_MAX_MEMCPY_LEN "mprq_max_memcpy_len"
61 
62 /* Device parameter to set the minimum number of Rx queues to enable MPRQ. */
63 #define MLX5_RXQS_MIN_MPRQ "rxqs_min_mprq"
64 
65 /* Device parameter to configure inline send. Deprecated, ignored.*/
66 #define MLX5_TXQ_INLINE "txq_inline"
67 
68 /* Device parameter to limit packet size to inline with ordinary SEND. */
69 #define MLX5_TXQ_INLINE_MAX "txq_inline_max"
70 
71 /* Device parameter to configure minimal data size to inline. */
72 #define MLX5_TXQ_INLINE_MIN "txq_inline_min"
73 
74 /* Device parameter to limit packet size to inline with Enhanced MPW. */
75 #define MLX5_TXQ_INLINE_MPW "txq_inline_mpw"
76 
77 /*
78  * Device parameter to configure the number of TX queues threshold for
79  * enabling inline send.
80  */
81 #define MLX5_TXQS_MIN_INLINE "txqs_min_inline"
82 
83 /*
84  * Device parameter to configure the number of TX queues threshold for
85  * enabling vectorized Tx, deprecated, ignored (no vectorized Tx routines).
86  */
87 #define MLX5_TXQS_MAX_VEC "txqs_max_vec"
88 
89 /* Device parameter to enable multi-packet send WQEs. */
90 #define MLX5_TXQ_MPW_EN "txq_mpw_en"
91 
92 /*
93  * Device parameter to force doorbell register mapping
94  * to non-cahed region eliminating the extra write memory barrier.
95  */
96 #define MLX5_TX_DB_NC "tx_db_nc"
97 
98 /*
99  * Device parameter to include 2 dsegs in the title WQEBB.
100  * Deprecated, ignored.
101  */
102 #define MLX5_TXQ_MPW_HDR_DSEG_EN "txq_mpw_hdr_dseg_en"
103 
104 /*
105  * Device parameter to limit the size of inlining packet.
106  * Deprecated, ignored.
107  */
108 #define MLX5_TXQ_MAX_INLINE_LEN "txq_max_inline_len"
109 
110 /*
111  * Device parameter to enable Tx scheduling on timestamps
112  * and specify the packet pacing granularity in nanoseconds.
113  */
114 #define MLX5_TX_PP "tx_pp"
115 
116 /*
117  * Device parameter to specify skew in nanoseconds on Tx datapath,
118  * it represents the time between SQ start WQE processing and
119  * appearing actual packet data on the wire.
120  */
121 #define MLX5_TX_SKEW "tx_skew"
122 
123 /*
124  * Device parameter to enable hardware Tx vector.
125  * Deprecated, ignored (no vectorized Tx routines anymore).
126  */
127 #define MLX5_TX_VEC_EN "tx_vec_en"
128 
129 /* Device parameter to enable hardware Rx vector. */
130 #define MLX5_RX_VEC_EN "rx_vec_en"
131 
132 /* Allow L3 VXLAN flow creation. */
133 #define MLX5_L3_VXLAN_EN "l3_vxlan_en"
134 
135 /* Activate DV E-Switch flow steering. */
136 #define MLX5_DV_ESW_EN "dv_esw_en"
137 
138 /* Activate DV flow steering. */
139 #define MLX5_DV_FLOW_EN "dv_flow_en"
140 
141 /* Enable extensive flow metadata support. */
142 #define MLX5_DV_XMETA_EN "dv_xmeta_en"
143 
144 /* Device parameter to let the user manage the lacp traffic of bonded device */
145 #define MLX5_LACP_BY_USER "lacp_by_user"
146 
147 /* Activate Netlink support in VF mode. */
148 #define MLX5_VF_NL_EN "vf_nl_en"
149 
150 /* Enable extending memsegs when creating a MR. */
151 #define MLX5_MR_EXT_MEMSEG_EN "mr_ext_memseg_en"
152 
153 /* Select port representors to instantiate. */
154 #define MLX5_REPRESENTOR "representor"
155 
156 /* Device parameter to configure the maximum number of dump files per queue. */
157 #define MLX5_MAX_DUMP_FILES_NUM "max_dump_files_num"
158 
159 /* Configure timeout of LRO session (in microseconds). */
160 #define MLX5_LRO_TIMEOUT_USEC "lro_timeout_usec"
161 
162 /*
163  * Device parameter to configure the total data buffer size for a single
164  * hairpin queue (logarithm value).
165  */
166 #define MLX5_HP_BUF_SIZE "hp_buf_log_sz"
167 
168 /* Flow memory reclaim mode. */
169 #define MLX5_RECLAIM_MEM "reclaim_mem_mode"
170 
171 /* The default memory allocator used in PMD. */
172 #define MLX5_SYS_MEM_EN "sys_mem_en"
173 /* Decap will be used or not. */
174 #define MLX5_DECAP_EN "decap_en"
175 
176 /* Shared memory between primary and secondary processes. */
177 struct mlx5_shared_data *mlx5_shared_data;
178 
179 /** Driver-specific log messages type. */
180 int mlx5_logtype;
181 
182 static LIST_HEAD(, mlx5_dev_ctx_shared) mlx5_dev_ctx_list =
183 						LIST_HEAD_INITIALIZER();
184 static pthread_mutex_t mlx5_dev_ctx_list_mutex;
185 static const struct mlx5_indexed_pool_config mlx5_ipool_cfg[] = {
186 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
187 	[MLX5_IPOOL_DECAP_ENCAP] = {
188 		.size = sizeof(struct mlx5_flow_dv_encap_decap_resource),
189 		.trunk_size = 64,
190 		.grow_trunk = 3,
191 		.grow_shift = 2,
192 		.need_lock = 1,
193 		.release_mem_en = 1,
194 		.malloc = mlx5_malloc,
195 		.free = mlx5_free,
196 		.type = "mlx5_encap_decap_ipool",
197 	},
198 	[MLX5_IPOOL_PUSH_VLAN] = {
199 		.size = sizeof(struct mlx5_flow_dv_push_vlan_action_resource),
200 		.trunk_size = 64,
201 		.grow_trunk = 3,
202 		.grow_shift = 2,
203 		.need_lock = 1,
204 		.release_mem_en = 1,
205 		.malloc = mlx5_malloc,
206 		.free = mlx5_free,
207 		.type = "mlx5_push_vlan_ipool",
208 	},
209 	[MLX5_IPOOL_TAG] = {
210 		.size = sizeof(struct mlx5_flow_dv_tag_resource),
211 		.trunk_size = 64,
212 		.grow_trunk = 3,
213 		.grow_shift = 2,
214 		.need_lock = 1,
215 		.release_mem_en = 1,
216 		.malloc = mlx5_malloc,
217 		.free = mlx5_free,
218 		.type = "mlx5_tag_ipool",
219 	},
220 	[MLX5_IPOOL_PORT_ID] = {
221 		.size = sizeof(struct mlx5_flow_dv_port_id_action_resource),
222 		.trunk_size = 64,
223 		.grow_trunk = 3,
224 		.grow_shift = 2,
225 		.need_lock = 1,
226 		.release_mem_en = 1,
227 		.malloc = mlx5_malloc,
228 		.free = mlx5_free,
229 		.type = "mlx5_port_id_ipool",
230 	},
231 	[MLX5_IPOOL_JUMP] = {
232 		.size = sizeof(struct mlx5_flow_tbl_data_entry),
233 		.trunk_size = 64,
234 		.grow_trunk = 3,
235 		.grow_shift = 2,
236 		.need_lock = 1,
237 		.release_mem_en = 1,
238 		.malloc = mlx5_malloc,
239 		.free = mlx5_free,
240 		.type = "mlx5_jump_ipool",
241 	},
242 	[MLX5_IPOOL_SAMPLE] = {
243 		.size = sizeof(struct mlx5_flow_dv_sample_resource),
244 		.trunk_size = 64,
245 		.grow_trunk = 3,
246 		.grow_shift = 2,
247 		.need_lock = 1,
248 		.release_mem_en = 1,
249 		.malloc = mlx5_malloc,
250 		.free = mlx5_free,
251 		.type = "mlx5_sample_ipool",
252 	},
253 	[MLX5_IPOOL_DEST_ARRAY] = {
254 		.size = sizeof(struct mlx5_flow_dv_dest_array_resource),
255 		.trunk_size = 64,
256 		.grow_trunk = 3,
257 		.grow_shift = 2,
258 		.need_lock = 1,
259 		.release_mem_en = 1,
260 		.malloc = mlx5_malloc,
261 		.free = mlx5_free,
262 		.type = "mlx5_dest_array_ipool",
263 	},
264 	[MLX5_IPOOL_TUNNEL_ID] = {
265 		.size = sizeof(struct mlx5_flow_tunnel),
266 		.trunk_size = MLX5_MAX_TUNNELS,
267 		.need_lock = 1,
268 		.release_mem_en = 1,
269 		.type = "mlx5_tunnel_offload",
270 	},
271 	[MLX5_IPOOL_TNL_TBL_ID] = {
272 		.size = 0,
273 		.need_lock = 1,
274 		.type = "mlx5_flow_tnl_tbl_ipool",
275 	},
276 #endif
277 	[MLX5_IPOOL_MTR] = {
278 		.size = sizeof(struct mlx5_flow_meter),
279 		.trunk_size = 64,
280 		.grow_trunk = 3,
281 		.grow_shift = 2,
282 		.need_lock = 1,
283 		.release_mem_en = 1,
284 		.malloc = mlx5_malloc,
285 		.free = mlx5_free,
286 		.type = "mlx5_meter_ipool",
287 	},
288 	[MLX5_IPOOL_MCP] = {
289 		.size = sizeof(struct mlx5_flow_mreg_copy_resource),
290 		.trunk_size = 64,
291 		.grow_trunk = 3,
292 		.grow_shift = 2,
293 		.need_lock = 1,
294 		.release_mem_en = 1,
295 		.malloc = mlx5_malloc,
296 		.free = mlx5_free,
297 		.type = "mlx5_mcp_ipool",
298 	},
299 	[MLX5_IPOOL_HRXQ] = {
300 		.size = (sizeof(struct mlx5_hrxq) + MLX5_RSS_HASH_KEY_LEN),
301 		.trunk_size = 64,
302 		.grow_trunk = 3,
303 		.grow_shift = 2,
304 		.need_lock = 1,
305 		.release_mem_en = 1,
306 		.malloc = mlx5_malloc,
307 		.free = mlx5_free,
308 		.type = "mlx5_hrxq_ipool",
309 	},
310 	[MLX5_IPOOL_MLX5_FLOW] = {
311 		/*
312 		 * MLX5_IPOOL_MLX5_FLOW size varies for DV and VERBS flows.
313 		 * It set in run time according to PCI function configuration.
314 		 */
315 		.size = 0,
316 		.trunk_size = 64,
317 		.grow_trunk = 3,
318 		.grow_shift = 2,
319 		.need_lock = 1,
320 		.release_mem_en = 1,
321 		.malloc = mlx5_malloc,
322 		.free = mlx5_free,
323 		.type = "mlx5_flow_handle_ipool",
324 	},
325 	[MLX5_IPOOL_RTE_FLOW] = {
326 		.size = sizeof(struct rte_flow),
327 		.trunk_size = 4096,
328 		.need_lock = 1,
329 		.release_mem_en = 1,
330 		.malloc = mlx5_malloc,
331 		.free = mlx5_free,
332 		.type = "rte_flow_ipool",
333 	},
334 	[MLX5_IPOOL_RSS_EXPANTION_FLOW_ID] = {
335 		.size = 0,
336 		.need_lock = 1,
337 		.type = "mlx5_flow_rss_id_ipool",
338 	},
339 	[MLX5_IPOOL_RSS_SHARED_ACTIONS] = {
340 		.size = sizeof(struct mlx5_shared_action_rss),
341 		.trunk_size = 64,
342 		.grow_trunk = 3,
343 		.grow_shift = 2,
344 		.need_lock = 1,
345 		.release_mem_en = 1,
346 		.malloc = mlx5_malloc,
347 		.free = mlx5_free,
348 		.type = "mlx5_shared_action_rss",
349 	},
350 };
351 
352 
353 #define MLX5_FLOW_MIN_ID_POOL_SIZE 512
354 #define MLX5_ID_GENERATION_ARRAY_FACTOR 16
355 
356 #define MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE 4096
357 
358 /**
359  * Decide whether representor ID is a HPF(host PF) port on BF2.
360  *
361  * @param dev
362  *   Pointer to Ethernet device structure.
363  *
364  * @return
365  *   Non-zero if HPF, otherwise 0.
366  */
367 bool
368 mlx5_is_hpf(struct rte_eth_dev *dev)
369 {
370 	struct mlx5_priv *priv = dev->data->dev_private;
371 	uint16_t repr = MLX5_REPRESENTOR_REPR(priv->representor_id);
372 	int type = MLX5_REPRESENTOR_TYPE(priv->representor_id);
373 
374 	return priv->representor != 0 && type == RTE_ETH_REPRESENTOR_VF &&
375 	       MLX5_REPRESENTOR_REPR(-1) == repr;
376 }
377 
378 /**
379  * Initialize the ASO aging management structure.
380  *
381  * @param[in] sh
382  *   Pointer to mlx5_dev_ctx_shared object to free
383  *
384  * @return
385  *   0 on success, a negative errno value otherwise and rte_errno is set.
386  */
387 int
388 mlx5_flow_aso_age_mng_init(struct mlx5_dev_ctx_shared *sh)
389 {
390 	int err;
391 
392 	if (sh->aso_age_mng)
393 		return 0;
394 	sh->aso_age_mng = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*sh->aso_age_mng),
395 				      RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
396 	if (!sh->aso_age_mng) {
397 		DRV_LOG(ERR, "aso_age_mng allocation was failed.");
398 		rte_errno = ENOMEM;
399 		return -ENOMEM;
400 	}
401 	err = mlx5_aso_queue_init(sh);
402 	if (err) {
403 		mlx5_free(sh->aso_age_mng);
404 		return -1;
405 	}
406 	rte_spinlock_init(&sh->aso_age_mng->resize_sl);
407 	rte_spinlock_init(&sh->aso_age_mng->free_sl);
408 	LIST_INIT(&sh->aso_age_mng->free);
409 	return 0;
410 }
411 
412 /**
413  * Close and release all the resources of the ASO aging management structure.
414  *
415  * @param[in] sh
416  *   Pointer to mlx5_dev_ctx_shared object to free.
417  */
418 static void
419 mlx5_flow_aso_age_mng_close(struct mlx5_dev_ctx_shared *sh)
420 {
421 	int i, j;
422 
423 	mlx5_aso_queue_stop(sh);
424 	mlx5_aso_queue_uninit(sh);
425 	if (sh->aso_age_mng->pools) {
426 		struct mlx5_aso_age_pool *pool;
427 
428 		for (i = 0; i < sh->aso_age_mng->next; ++i) {
429 			pool = sh->aso_age_mng->pools[i];
430 			claim_zero(mlx5_devx_cmd_destroy
431 						(pool->flow_hit_aso_obj));
432 			for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j)
433 				if (pool->actions[j].dr_action)
434 					claim_zero
435 					    (mlx5_flow_os_destroy_flow_action
436 					      (pool->actions[j].dr_action));
437 			mlx5_free(pool);
438 		}
439 		mlx5_free(sh->aso_age_mng->pools);
440 	}
441 	mlx5_free(sh->aso_age_mng);
442 }
443 
444 /**
445  * Initialize the shared aging list information per port.
446  *
447  * @param[in] sh
448  *   Pointer to mlx5_dev_ctx_shared object.
449  */
450 static void
451 mlx5_flow_aging_init(struct mlx5_dev_ctx_shared *sh)
452 {
453 	uint32_t i;
454 	struct mlx5_age_info *age_info;
455 
456 	for (i = 0; i < sh->max_port; i++) {
457 		age_info = &sh->port[i].age_info;
458 		age_info->flags = 0;
459 		TAILQ_INIT(&age_info->aged_counters);
460 		LIST_INIT(&age_info->aged_aso);
461 		rte_spinlock_init(&age_info->aged_sl);
462 		MLX5_AGE_SET(age_info, MLX5_AGE_TRIGGER);
463 	}
464 }
465 
466 /**
467  * Initialize the counters management structure.
468  *
469  * @param[in] sh
470  *   Pointer to mlx5_dev_ctx_shared object to free
471  */
472 static void
473 mlx5_flow_counters_mng_init(struct mlx5_dev_ctx_shared *sh)
474 {
475 	int i;
476 
477 	memset(&sh->cmng, 0, sizeof(sh->cmng));
478 	TAILQ_INIT(&sh->cmng.flow_counters);
479 	sh->cmng.min_id = MLX5_CNT_BATCH_OFFSET;
480 	sh->cmng.max_id = -1;
481 	sh->cmng.last_pool_idx = POOL_IDX_INVALID;
482 	rte_spinlock_init(&sh->cmng.pool_update_sl);
483 	for (i = 0; i < MLX5_COUNTER_TYPE_MAX; i++) {
484 		TAILQ_INIT(&sh->cmng.counters[i]);
485 		rte_spinlock_init(&sh->cmng.csl[i]);
486 	}
487 }
488 
489 /**
490  * Destroy all the resources allocated for a counter memory management.
491  *
492  * @param[in] mng
493  *   Pointer to the memory management structure.
494  */
495 static void
496 mlx5_flow_destroy_counter_stat_mem_mng(struct mlx5_counter_stats_mem_mng *mng)
497 {
498 	uint8_t *mem = (uint8_t *)(uintptr_t)mng->raws[0].data;
499 
500 	LIST_REMOVE(mng, next);
501 	claim_zero(mlx5_devx_cmd_destroy(mng->dm));
502 	claim_zero(mlx5_os_umem_dereg(mng->umem));
503 	mlx5_free(mem);
504 }
505 
506 /**
507  * Close and release all the resources of the counters management.
508  *
509  * @param[in] sh
510  *   Pointer to mlx5_dev_ctx_shared object to free.
511  */
512 static void
513 mlx5_flow_counters_mng_close(struct mlx5_dev_ctx_shared *sh)
514 {
515 	struct mlx5_counter_stats_mem_mng *mng;
516 	int i, j;
517 	int retries = 1024;
518 
519 	rte_errno = 0;
520 	while (--retries) {
521 		rte_eal_alarm_cancel(mlx5_flow_query_alarm, sh);
522 		if (rte_errno != EINPROGRESS)
523 			break;
524 		rte_pause();
525 	}
526 
527 	if (sh->cmng.pools) {
528 		struct mlx5_flow_counter_pool *pool;
529 		uint16_t n_valid = sh->cmng.n_valid;
530 		bool fallback = sh->cmng.counter_fallback;
531 
532 		for (i = 0; i < n_valid; ++i) {
533 			pool = sh->cmng.pools[i];
534 			if (!fallback && pool->min_dcs)
535 				claim_zero(mlx5_devx_cmd_destroy
536 							       (pool->min_dcs));
537 			for (j = 0; j < MLX5_COUNTERS_PER_POOL; ++j) {
538 				struct mlx5_flow_counter *cnt =
539 						MLX5_POOL_GET_CNT(pool, j);
540 
541 				if (cnt->action)
542 					claim_zero
543 					 (mlx5_flow_os_destroy_flow_action
544 					  (cnt->action));
545 				if (fallback && MLX5_POOL_GET_CNT
546 				    (pool, j)->dcs_when_free)
547 					claim_zero(mlx5_devx_cmd_destroy
548 						   (cnt->dcs_when_free));
549 			}
550 			mlx5_free(pool);
551 		}
552 		mlx5_free(sh->cmng.pools);
553 	}
554 	mng = LIST_FIRST(&sh->cmng.mem_mngs);
555 	while (mng) {
556 		mlx5_flow_destroy_counter_stat_mem_mng(mng);
557 		mng = LIST_FIRST(&sh->cmng.mem_mngs);
558 	}
559 	memset(&sh->cmng, 0, sizeof(sh->cmng));
560 }
561 
562 /* Send FLOW_AGED event if needed. */
563 void
564 mlx5_age_event_prepare(struct mlx5_dev_ctx_shared *sh)
565 {
566 	struct mlx5_age_info *age_info;
567 	uint32_t i;
568 
569 	for (i = 0; i < sh->max_port; i++) {
570 		age_info = &sh->port[i].age_info;
571 		if (!MLX5_AGE_GET(age_info, MLX5_AGE_EVENT_NEW))
572 			continue;
573 		if (MLX5_AGE_GET(age_info, MLX5_AGE_TRIGGER))
574 			rte_eth_dev_callback_process
575 				(&rte_eth_devices[sh->port[i].devx_ih_port_id],
576 				RTE_ETH_EVENT_FLOW_AGED, NULL);
577 		age_info->flags = 0;
578 	}
579 }
580 
581 /**
582  * Initialize the flow resources' indexed mempool.
583  *
584  * @param[in] sh
585  *   Pointer to mlx5_dev_ctx_shared object.
586  * @param[in] sh
587  *   Pointer to user dev config.
588  */
589 static void
590 mlx5_flow_ipool_create(struct mlx5_dev_ctx_shared *sh,
591 		       const struct mlx5_dev_config *config)
592 {
593 	uint8_t i;
594 	struct mlx5_indexed_pool_config cfg;
595 
596 	for (i = 0; i < MLX5_IPOOL_MAX; ++i) {
597 		cfg = mlx5_ipool_cfg[i];
598 		switch (i) {
599 		default:
600 			break;
601 		/*
602 		 * Set MLX5_IPOOL_MLX5_FLOW ipool size
603 		 * according to PCI function flow configuration.
604 		 */
605 		case MLX5_IPOOL_MLX5_FLOW:
606 			cfg.size = config->dv_flow_en ?
607 				sizeof(struct mlx5_flow_handle) :
608 				MLX5_FLOW_HANDLE_VERBS_SIZE;
609 			break;
610 		}
611 		if (config->reclaim_mode)
612 			cfg.release_mem_en = 1;
613 		sh->ipool[i] = mlx5_ipool_create(&cfg);
614 	}
615 }
616 
617 /**
618  * Release the flow resources' indexed mempool.
619  *
620  * @param[in] sh
621  *   Pointer to mlx5_dev_ctx_shared object.
622  */
623 static void
624 mlx5_flow_ipool_destroy(struct mlx5_dev_ctx_shared *sh)
625 {
626 	uint8_t i;
627 
628 	for (i = 0; i < MLX5_IPOOL_MAX; ++i)
629 		mlx5_ipool_destroy(sh->ipool[i]);
630 }
631 
632 /*
633  * Check if dynamic flex parser for eCPRI already exists.
634  *
635  * @param dev
636  *   Pointer to Ethernet device structure.
637  *
638  * @return
639  *   true on exists, false on not.
640  */
641 bool
642 mlx5_flex_parser_ecpri_exist(struct rte_eth_dev *dev)
643 {
644 	struct mlx5_priv *priv = dev->data->dev_private;
645 	struct mlx5_flex_parser_profiles *prf =
646 				&priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
647 
648 	return !!prf->obj;
649 }
650 
651 /*
652  * Allocation of a flex parser for eCPRI. Once created, this parser related
653  * resources will be held until the device is closed.
654  *
655  * @param dev
656  *   Pointer to Ethernet device structure.
657  *
658  * @return
659  *   0 on success, a negative errno value otherwise and rte_errno is set.
660  */
661 int
662 mlx5_flex_parser_ecpri_alloc(struct rte_eth_dev *dev)
663 {
664 	struct mlx5_priv *priv = dev->data->dev_private;
665 	struct mlx5_flex_parser_profiles *prf =
666 				&priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
667 	struct mlx5_devx_graph_node_attr node = {
668 		.modify_field_select = 0,
669 	};
670 	uint32_t ids[8];
671 	int ret;
672 
673 	if (!priv->config.hca_attr.parse_graph_flex_node) {
674 		DRV_LOG(ERR, "Dynamic flex parser is not supported "
675 			"for device %s.", priv->dev_data->name);
676 		return -ENOTSUP;
677 	}
678 	node.header_length_mode = MLX5_GRAPH_NODE_LEN_FIXED;
679 	/* 8 bytes now: 4B common header + 4B message body header. */
680 	node.header_length_base_value = 0x8;
681 	/* After MAC layer: Ether / VLAN. */
682 	node.in[0].arc_parse_graph_node = MLX5_GRAPH_ARC_NODE_MAC;
683 	/* Type of compared condition should be 0xAEFE in the L2 layer. */
684 	node.in[0].compare_condition_value = RTE_ETHER_TYPE_ECPRI;
685 	/* Sample #0: type in common header. */
686 	node.sample[0].flow_match_sample_en = 1;
687 	/* Fixed offset. */
688 	node.sample[0].flow_match_sample_offset_mode = 0x0;
689 	/* Only the 2nd byte will be used. */
690 	node.sample[0].flow_match_sample_field_base_offset = 0x0;
691 	/* Sample #1: message payload. */
692 	node.sample[1].flow_match_sample_en = 1;
693 	/* Fixed offset. */
694 	node.sample[1].flow_match_sample_offset_mode = 0x0;
695 	/*
696 	 * Only the first two bytes will be used right now, and its offset will
697 	 * start after the common header that with the length of a DW(u32).
698 	 */
699 	node.sample[1].flow_match_sample_field_base_offset = sizeof(uint32_t);
700 	prf->obj = mlx5_devx_cmd_create_flex_parser(priv->sh->ctx, &node);
701 	if (!prf->obj) {
702 		DRV_LOG(ERR, "Failed to create flex parser node object.");
703 		return (rte_errno == 0) ? -ENODEV : -rte_errno;
704 	}
705 	prf->num = 2;
706 	ret = mlx5_devx_cmd_query_parse_samples(prf->obj, ids, prf->num);
707 	if (ret) {
708 		DRV_LOG(ERR, "Failed to query sample IDs.");
709 		return (rte_errno == 0) ? -ENODEV : -rte_errno;
710 	}
711 	prf->offset[0] = 0x0;
712 	prf->offset[1] = sizeof(uint32_t);
713 	prf->ids[0] = ids[0];
714 	prf->ids[1] = ids[1];
715 	return 0;
716 }
717 
718 /*
719  * Destroy the flex parser node, including the parser itself, input / output
720  * arcs and DW samples. Resources could be reused then.
721  *
722  * @param dev
723  *   Pointer to Ethernet device structure.
724  */
725 static void
726 mlx5_flex_parser_ecpri_release(struct rte_eth_dev *dev)
727 {
728 	struct mlx5_priv *priv = dev->data->dev_private;
729 	struct mlx5_flex_parser_profiles *prf =
730 				&priv->sh->fp[MLX5_FLEX_PARSER_ECPRI_0];
731 
732 	if (prf->obj)
733 		mlx5_devx_cmd_destroy(prf->obj);
734 	prf->obj = NULL;
735 }
736 
737 /*
738  * Allocate Rx and Tx UARs in robust fashion.
739  * This routine handles the following UAR allocation issues:
740  *
741  *  - tries to allocate the UAR with the most appropriate memory
742  *    mapping type from the ones supported by the host
743  *
744  *  - tries to allocate the UAR with non-NULL base address
745  *    OFED 5.0.x and Upstream rdma_core before v29 returned the NULL as
746  *    UAR base address if UAR was not the first object in the UAR page.
747  *    It caused the PMD failure and we should try to get another UAR
748  *    till we get the first one with non-NULL base address returned.
749  */
750 static int
751 mlx5_alloc_rxtx_uars(struct mlx5_dev_ctx_shared *sh,
752 		     const struct mlx5_dev_config *config)
753 {
754 	uint32_t uar_mapping, retry;
755 	int err = 0;
756 	void *base_addr;
757 
758 	for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
759 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
760 		/* Control the mapping type according to the settings. */
761 		uar_mapping = (config->dbnc == MLX5_TXDB_NCACHED) ?
762 			      MLX5DV_UAR_ALLOC_TYPE_NC :
763 			      MLX5DV_UAR_ALLOC_TYPE_BF;
764 #else
765 		RTE_SET_USED(config);
766 		/*
767 		 * It seems we have no way to control the memory mapping type
768 		 * for the UAR, the default "Write-Combining" type is supposed.
769 		 * The UAR initialization on queue creation queries the
770 		 * actual mapping type done by Verbs/kernel and setups the
771 		 * PMD datapath accordingly.
772 		 */
773 		uar_mapping = 0;
774 #endif
775 		sh->tx_uar = mlx5_glue->devx_alloc_uar(sh->ctx, uar_mapping);
776 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
777 		if (!sh->tx_uar &&
778 		    uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
779 			if (config->dbnc == MLX5_TXDB_CACHED ||
780 			    config->dbnc == MLX5_TXDB_HEURISTIC)
781 				DRV_LOG(WARNING, "Devarg tx_db_nc setting "
782 						 "is not supported by DevX");
783 			/*
784 			 * In some environments like virtual machine
785 			 * the Write Combining mapped might be not supported
786 			 * and UAR allocation fails. We try "Non-Cached"
787 			 * mapping for the case. The tx_burst routines take
788 			 * the UAR mapping type into account on UAR setup
789 			 * on queue creation.
790 			 */
791 			DRV_LOG(DEBUG, "Failed to allocate Tx DevX UAR (BF)");
792 			uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
793 			sh->tx_uar = mlx5_glue->devx_alloc_uar
794 							(sh->ctx, uar_mapping);
795 		} else if (!sh->tx_uar &&
796 			   uar_mapping == MLX5DV_UAR_ALLOC_TYPE_NC) {
797 			if (config->dbnc == MLX5_TXDB_NCACHED)
798 				DRV_LOG(WARNING, "Devarg tx_db_nc settings "
799 						 "is not supported by DevX");
800 			/*
801 			 * If Verbs/kernel does not support "Non-Cached"
802 			 * try the "Write-Combining".
803 			 */
804 			DRV_LOG(DEBUG, "Failed to allocate Tx DevX UAR (NC)");
805 			uar_mapping = MLX5DV_UAR_ALLOC_TYPE_BF;
806 			sh->tx_uar = mlx5_glue->devx_alloc_uar
807 							(sh->ctx, uar_mapping);
808 		}
809 #endif
810 		if (!sh->tx_uar) {
811 			DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (BF/NC)");
812 			err = ENOMEM;
813 			goto exit;
814 		}
815 		base_addr = mlx5_os_get_devx_uar_base_addr(sh->tx_uar);
816 		if (base_addr)
817 			break;
818 		/*
819 		 * The UARs are allocated by rdma_core within the
820 		 * IB device context, on context closure all UARs
821 		 * will be freed, should be no memory/object leakage.
822 		 */
823 		DRV_LOG(DEBUG, "Retrying to allocate Tx DevX UAR");
824 		sh->tx_uar = NULL;
825 	}
826 	/* Check whether we finally succeeded with valid UAR allocation. */
827 	if (!sh->tx_uar) {
828 		DRV_LOG(ERR, "Failed to allocate Tx DevX UAR (NULL base)");
829 		err = ENOMEM;
830 		goto exit;
831 	}
832 	for (retry = 0; retry < MLX5_ALLOC_UAR_RETRY; ++retry) {
833 		uar_mapping = 0;
834 		sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
835 							(sh->ctx, uar_mapping);
836 #ifdef MLX5DV_UAR_ALLOC_TYPE_NC
837 		if (!sh->devx_rx_uar &&
838 		    uar_mapping == MLX5DV_UAR_ALLOC_TYPE_BF) {
839 			/*
840 			 * Rx UAR is used to control interrupts only,
841 			 * should be no datapath noticeable impact,
842 			 * can try "Non-Cached" mapping safely.
843 			 */
844 			DRV_LOG(DEBUG, "Failed to allocate Rx DevX UAR (BF)");
845 			uar_mapping = MLX5DV_UAR_ALLOC_TYPE_NC;
846 			sh->devx_rx_uar = mlx5_glue->devx_alloc_uar
847 							(sh->ctx, uar_mapping);
848 		}
849 #endif
850 		if (!sh->devx_rx_uar) {
851 			DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (BF/NC)");
852 			err = ENOMEM;
853 			goto exit;
854 		}
855 		base_addr = mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar);
856 		if (base_addr)
857 			break;
858 		/*
859 		 * The UARs are allocated by rdma_core within the
860 		 * IB device context, on context closure all UARs
861 		 * will be freed, should be no memory/object leakage.
862 		 */
863 		DRV_LOG(DEBUG, "Retrying to allocate Rx DevX UAR");
864 		sh->devx_rx_uar = NULL;
865 	}
866 	/* Check whether we finally succeeded with valid UAR allocation. */
867 	if (!sh->devx_rx_uar) {
868 		DRV_LOG(ERR, "Failed to allocate Rx DevX UAR (NULL base)");
869 		err = ENOMEM;
870 	}
871 exit:
872 	return err;
873 }
874 
875 /**
876  * Allocate shared device context. If there is multiport device the
877  * master and representors will share this context, if there is single
878  * port dedicated device, the context will be used by only given
879  * port due to unification.
880  *
881  * Routine first searches the context for the specified device name,
882  * if found the shared context assumed and reference counter is incremented.
883  * If no context found the new one is created and initialized with specified
884  * device context and parameters.
885  *
886  * @param[in] spawn
887  *   Pointer to the device attributes (name, port, etc).
888  * @param[in] config
889  *   Pointer to device configuration structure.
890  *
891  * @return
892  *   Pointer to mlx5_dev_ctx_shared object on success,
893  *   otherwise NULL and rte_errno is set.
894  */
895 struct mlx5_dev_ctx_shared *
896 mlx5_alloc_shared_dev_ctx(const struct mlx5_dev_spawn_data *spawn,
897 			   const struct mlx5_dev_config *config)
898 {
899 	struct mlx5_dev_ctx_shared *sh;
900 	int err = 0;
901 	uint32_t i;
902 	struct mlx5_devx_tis_attr tis_attr = { 0 };
903 
904 	MLX5_ASSERT(spawn);
905 	/* Secondary process should not create the shared context. */
906 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
907 	pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
908 	/* Search for IB context by device name. */
909 	LIST_FOREACH(sh, &mlx5_dev_ctx_list, next) {
910 		if (!strcmp(sh->ibdev_name,
911 			mlx5_os_get_dev_device_name(spawn->phys_dev))) {
912 			sh->refcnt++;
913 			goto exit;
914 		}
915 	}
916 	/* No device found, we have to create new shared context. */
917 	MLX5_ASSERT(spawn->max_port);
918 	sh = mlx5_malloc(MLX5_MEM_ZERO | MLX5_MEM_RTE,
919 			 sizeof(struct mlx5_dev_ctx_shared) +
920 			 spawn->max_port *
921 			 sizeof(struct mlx5_dev_shared_port),
922 			 RTE_CACHE_LINE_SIZE, SOCKET_ID_ANY);
923 	if (!sh) {
924 		DRV_LOG(ERR, "shared context allocation failure");
925 		rte_errno  = ENOMEM;
926 		goto exit;
927 	}
928 	if (spawn->bond_info)
929 		sh->bond = *spawn->bond_info;
930 	err = mlx5_os_open_device(spawn, config, sh);
931 	if (!sh->ctx)
932 		goto error;
933 	err = mlx5_os_get_dev_attr(sh->ctx, &sh->device_attr);
934 	if (err) {
935 		DRV_LOG(DEBUG, "mlx5_os_get_dev_attr() failed");
936 		goto error;
937 	}
938 	sh->refcnt = 1;
939 	sh->max_port = spawn->max_port;
940 	strncpy(sh->ibdev_name, mlx5_os_get_ctx_device_name(sh->ctx),
941 		sizeof(sh->ibdev_name) - 1);
942 	strncpy(sh->ibdev_path, mlx5_os_get_ctx_device_path(sh->ctx),
943 		sizeof(sh->ibdev_path) - 1);
944 	/*
945 	 * Setting port_id to max unallowed value means
946 	 * there is no interrupt subhandler installed for
947 	 * the given port index i.
948 	 */
949 	for (i = 0; i < sh->max_port; i++) {
950 		sh->port[i].ih_port_id = RTE_MAX_ETHPORTS;
951 		sh->port[i].devx_ih_port_id = RTE_MAX_ETHPORTS;
952 	}
953 	sh->pd = mlx5_os_alloc_pd(sh->ctx);
954 	if (sh->pd == NULL) {
955 		DRV_LOG(ERR, "PD allocation failure");
956 		err = ENOMEM;
957 		goto error;
958 	}
959 	if (sh->devx) {
960 		err = mlx5_os_get_pdn(sh->pd, &sh->pdn);
961 		if (err) {
962 			DRV_LOG(ERR, "Fail to extract pdn from PD");
963 			goto error;
964 		}
965 		sh->td = mlx5_devx_cmd_create_td(sh->ctx);
966 		if (!sh->td) {
967 			DRV_LOG(ERR, "TD allocation failure");
968 			err = ENOMEM;
969 			goto error;
970 		}
971 		tis_attr.transport_domain = sh->td->id;
972 		sh->tis = mlx5_devx_cmd_create_tis(sh->ctx, &tis_attr);
973 		if (!sh->tis) {
974 			DRV_LOG(ERR, "TIS allocation failure");
975 			err = ENOMEM;
976 			goto error;
977 		}
978 		err = mlx5_alloc_rxtx_uars(sh, config);
979 		if (err)
980 			goto error;
981 		MLX5_ASSERT(sh->tx_uar);
982 		MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->tx_uar));
983 
984 		MLX5_ASSERT(sh->devx_rx_uar);
985 		MLX5_ASSERT(mlx5_os_get_devx_uar_base_addr(sh->devx_rx_uar));
986 	}
987 #ifndef RTE_ARCH_64
988 	/* Initialize UAR access locks for 32bit implementations. */
989 	rte_spinlock_init(&sh->uar_lock_cq);
990 	for (i = 0; i < MLX5_UAR_PAGE_NUM_MAX; i++)
991 		rte_spinlock_init(&sh->uar_lock[i]);
992 #endif
993 	/*
994 	 * Once the device is added to the list of memory event
995 	 * callback, its global MR cache table cannot be expanded
996 	 * on the fly because of deadlock. If it overflows, lookup
997 	 * should be done by searching MR list linearly, which is slow.
998 	 *
999 	 * At this point the device is not added to the memory
1000 	 * event list yet, context is just being created.
1001 	 */
1002 	err = mlx5_mr_btree_init(&sh->share_cache.cache,
1003 				 MLX5_MR_BTREE_CACHE_N * 2,
1004 				 spawn->pci_dev->device.numa_node);
1005 	if (err) {
1006 		err = rte_errno;
1007 		goto error;
1008 	}
1009 	mlx5_os_set_reg_mr_cb(&sh->share_cache.reg_mr_cb,
1010 			      &sh->share_cache.dereg_mr_cb);
1011 	mlx5_os_dev_shared_handler_install(sh);
1012 	sh->cnt_id_tbl = mlx5_l3t_create(MLX5_L3T_TYPE_DWORD);
1013 	if (!sh->cnt_id_tbl) {
1014 		err = rte_errno;
1015 		goto error;
1016 	}
1017 	if (LIST_EMPTY(&mlx5_dev_ctx_list)) {
1018 		err = mlx5_flow_os_init_workspace_once();
1019 		if (err)
1020 			goto error;
1021 	}
1022 	mlx5_flow_aging_init(sh);
1023 	mlx5_flow_counters_mng_init(sh);
1024 	mlx5_flow_ipool_create(sh, config);
1025 	/* Add device to memory callback list. */
1026 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1027 	LIST_INSERT_HEAD(&mlx5_shared_data->mem_event_cb_list,
1028 			 sh, mem_event_cb);
1029 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1030 	/* Add context to the global device list. */
1031 	LIST_INSERT_HEAD(&mlx5_dev_ctx_list, sh, next);
1032 	rte_spinlock_init(&sh->geneve_tlv_opt_sl);
1033 exit:
1034 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1035 	return sh;
1036 error:
1037 	pthread_mutex_destroy(&sh->txpp.mutex);
1038 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1039 	MLX5_ASSERT(sh);
1040 	if (sh->cnt_id_tbl)
1041 		mlx5_l3t_destroy(sh->cnt_id_tbl);
1042 	if (sh->tis)
1043 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1044 	if (sh->td)
1045 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
1046 	if (sh->devx_rx_uar)
1047 		mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1048 	if (sh->tx_uar)
1049 		mlx5_glue->devx_free_uar(sh->tx_uar);
1050 	if (sh->pd)
1051 		claim_zero(mlx5_os_dealloc_pd(sh->pd));
1052 	if (sh->ctx)
1053 		claim_zero(mlx5_glue->close_device(sh->ctx));
1054 	mlx5_free(sh);
1055 	MLX5_ASSERT(err > 0);
1056 	rte_errno = err;
1057 	return NULL;
1058 }
1059 
1060 /**
1061  * Free shared IB device context. Decrement counter and if zero free
1062  * all allocated resources and close handles.
1063  *
1064  * @param[in] sh
1065  *   Pointer to mlx5_dev_ctx_shared object to free
1066  */
1067 void
1068 mlx5_free_shared_dev_ctx(struct mlx5_dev_ctx_shared *sh)
1069 {
1070 	pthread_mutex_lock(&mlx5_dev_ctx_list_mutex);
1071 #ifdef RTE_LIBRTE_MLX5_DEBUG
1072 	/* Check the object presence in the list. */
1073 	struct mlx5_dev_ctx_shared *lctx;
1074 
1075 	LIST_FOREACH(lctx, &mlx5_dev_ctx_list, next)
1076 		if (lctx == sh)
1077 			break;
1078 	MLX5_ASSERT(lctx);
1079 	if (lctx != sh) {
1080 		DRV_LOG(ERR, "Freeing non-existing shared IB context");
1081 		goto exit;
1082 	}
1083 #endif
1084 	MLX5_ASSERT(sh);
1085 	MLX5_ASSERT(sh->refcnt);
1086 	/* Secondary process should not free the shared context. */
1087 	MLX5_ASSERT(rte_eal_process_type() == RTE_PROC_PRIMARY);
1088 	if (--sh->refcnt)
1089 		goto exit;
1090 	/* Remove from memory callback device list. */
1091 	rte_rwlock_write_lock(&mlx5_shared_data->mem_event_rwlock);
1092 	LIST_REMOVE(sh, mem_event_cb);
1093 	rte_rwlock_write_unlock(&mlx5_shared_data->mem_event_rwlock);
1094 	/* Release created Memory Regions. */
1095 	mlx5_mr_release_cache(&sh->share_cache);
1096 	/* Remove context from the global device list. */
1097 	LIST_REMOVE(sh, next);
1098 	/* Release flow workspaces objects on the last device. */
1099 	if (LIST_EMPTY(&mlx5_dev_ctx_list))
1100 		mlx5_flow_os_release_workspace();
1101 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1102 	/*
1103 	 *  Ensure there is no async event handler installed.
1104 	 *  Only primary process handles async device events.
1105 	 **/
1106 	mlx5_flow_counters_mng_close(sh);
1107 	if (sh->aso_age_mng) {
1108 		mlx5_flow_aso_age_mng_close(sh);
1109 		sh->aso_age_mng = NULL;
1110 	}
1111 	mlx5_flow_ipool_destroy(sh);
1112 	mlx5_os_dev_shared_handler_uninstall(sh);
1113 	if (sh->cnt_id_tbl) {
1114 		mlx5_l3t_destroy(sh->cnt_id_tbl);
1115 		sh->cnt_id_tbl = NULL;
1116 	}
1117 	if (sh->tx_uar) {
1118 		mlx5_glue->devx_free_uar(sh->tx_uar);
1119 		sh->tx_uar = NULL;
1120 	}
1121 	if (sh->pd)
1122 		claim_zero(mlx5_os_dealloc_pd(sh->pd));
1123 	if (sh->tis)
1124 		claim_zero(mlx5_devx_cmd_destroy(sh->tis));
1125 	if (sh->td)
1126 		claim_zero(mlx5_devx_cmd_destroy(sh->td));
1127 	if (sh->devx_rx_uar)
1128 		mlx5_glue->devx_free_uar(sh->devx_rx_uar);
1129 	if (sh->ctx)
1130 		claim_zero(mlx5_glue->close_device(sh->ctx));
1131 	MLX5_ASSERT(sh->geneve_tlv_option_resource == NULL);
1132 	pthread_mutex_destroy(&sh->txpp.mutex);
1133 	mlx5_free(sh);
1134 	return;
1135 exit:
1136 	pthread_mutex_unlock(&mlx5_dev_ctx_list_mutex);
1137 }
1138 
1139 /**
1140  * Destroy table hash list.
1141  *
1142  * @param[in] priv
1143  *   Pointer to the private device data structure.
1144  */
1145 void
1146 mlx5_free_table_hash_list(struct mlx5_priv *priv)
1147 {
1148 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1149 
1150 	if (!sh->flow_tbls)
1151 		return;
1152 	mlx5_hlist_destroy(sh->flow_tbls);
1153 }
1154 
1155 /**
1156  * Initialize flow table hash list and create the root tables entry
1157  * for each domain.
1158  *
1159  * @param[in] priv
1160  *   Pointer to the private device data structure.
1161  *
1162  * @return
1163  *   Zero on success, positive error code otherwise.
1164  */
1165 int
1166 mlx5_alloc_table_hash_list(struct mlx5_priv *priv __rte_unused)
1167 {
1168 	int err = 0;
1169 	/* Tables are only used in DV and DR modes. */
1170 #if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
1171 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1172 	char s[MLX5_HLIST_NAMESIZE];
1173 
1174 	MLX5_ASSERT(sh);
1175 	snprintf(s, sizeof(s), "%s_flow_table", priv->sh->ibdev_name);
1176 	sh->flow_tbls = mlx5_hlist_create(s, MLX5_FLOW_TABLE_HLIST_ARRAY_SIZE,
1177 					  0, 0, flow_dv_tbl_create_cb,
1178 					  flow_dv_tbl_match_cb,
1179 					  flow_dv_tbl_remove_cb);
1180 	if (!sh->flow_tbls) {
1181 		DRV_LOG(ERR, "flow tables with hash creation failed.");
1182 		err = ENOMEM;
1183 		return err;
1184 	}
1185 	sh->flow_tbls->ctx = sh;
1186 #ifndef HAVE_MLX5DV_DR
1187 	struct rte_flow_error error;
1188 	struct rte_eth_dev *dev = &rte_eth_devices[priv->dev_data->port_id];
1189 
1190 	/*
1191 	 * In case we have not DR support, the zero tables should be created
1192 	 * because DV expect to see them even if they cannot be created by
1193 	 * RDMA-CORE.
1194 	 */
1195 	if (!flow_dv_tbl_resource_get(dev, 0, 0, 0, 0, NULL, 0, 1, &error) ||
1196 	    !flow_dv_tbl_resource_get(dev, 0, 1, 0, 0, NULL, 0, 1, &error) ||
1197 	    !flow_dv_tbl_resource_get(dev, 0, 0, 1, 0, NULL, 0, 1, &error)) {
1198 		err = ENOMEM;
1199 		goto error;
1200 	}
1201 	return err;
1202 error:
1203 	mlx5_free_table_hash_list(priv);
1204 #endif /* HAVE_MLX5DV_DR */
1205 #endif
1206 	return err;
1207 }
1208 
1209 /**
1210  * Retrieve integer value from environment variable.
1211  *
1212  * @param[in] name
1213  *   Environment variable name.
1214  *
1215  * @return
1216  *   Integer value, 0 if the variable is not set.
1217  */
1218 int
1219 mlx5_getenv_int(const char *name)
1220 {
1221 	const char *val = getenv(name);
1222 
1223 	if (val == NULL)
1224 		return 0;
1225 	return atoi(val);
1226 }
1227 
1228 /**
1229  * DPDK callback to add udp tunnel port
1230  *
1231  * @param[in] dev
1232  *   A pointer to eth_dev
1233  * @param[in] udp_tunnel
1234  *   A pointer to udp tunnel
1235  *
1236  * @return
1237  *   0 on valid udp ports and tunnels, -ENOTSUP otherwise.
1238  */
1239 int
1240 mlx5_udp_tunnel_port_add(struct rte_eth_dev *dev __rte_unused,
1241 			 struct rte_eth_udp_tunnel *udp_tunnel)
1242 {
1243 	MLX5_ASSERT(udp_tunnel != NULL);
1244 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN &&
1245 	    udp_tunnel->udp_port == 4789)
1246 		return 0;
1247 	if (udp_tunnel->prot_type == RTE_TUNNEL_TYPE_VXLAN_GPE &&
1248 	    udp_tunnel->udp_port == 4790)
1249 		return 0;
1250 	return -ENOTSUP;
1251 }
1252 
1253 /**
1254  * Initialize process private data structure.
1255  *
1256  * @param dev
1257  *   Pointer to Ethernet device structure.
1258  *
1259  * @return
1260  *   0 on success, a negative errno value otherwise and rte_errno is set.
1261  */
1262 int
1263 mlx5_proc_priv_init(struct rte_eth_dev *dev)
1264 {
1265 	struct mlx5_priv *priv = dev->data->dev_private;
1266 	struct mlx5_proc_priv *ppriv;
1267 	size_t ppriv_size;
1268 
1269 	/*
1270 	 * UAR register table follows the process private structure. BlueFlame
1271 	 * registers for Tx queues are stored in the table.
1272 	 */
1273 	ppriv_size =
1274 		sizeof(struct mlx5_proc_priv) + priv->txqs_n * sizeof(void *);
1275 	ppriv = mlx5_malloc(MLX5_MEM_RTE | MLX5_MEM_ZERO, ppriv_size,
1276 			    RTE_CACHE_LINE_SIZE, dev->device->numa_node);
1277 	if (!ppriv) {
1278 		rte_errno = ENOMEM;
1279 		return -rte_errno;
1280 	}
1281 	ppriv->uar_table_sz = priv->txqs_n;
1282 	dev->process_private = ppriv;
1283 	return 0;
1284 }
1285 
1286 /**
1287  * Un-initialize process private data structure.
1288  *
1289  * @param dev
1290  *   Pointer to Ethernet device structure.
1291  */
1292 void
1293 mlx5_proc_priv_uninit(struct rte_eth_dev *dev)
1294 {
1295 	if (!dev->process_private)
1296 		return;
1297 	mlx5_free(dev->process_private);
1298 	dev->process_private = NULL;
1299 }
1300 
1301 /**
1302  * DPDK callback to close the device.
1303  *
1304  * Destroy all queues and objects, free memory.
1305  *
1306  * @param dev
1307  *   Pointer to Ethernet device structure.
1308  */
1309 int
1310 mlx5_dev_close(struct rte_eth_dev *dev)
1311 {
1312 	struct mlx5_priv *priv = dev->data->dev_private;
1313 	unsigned int i;
1314 	int ret;
1315 
1316 	if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1317 		/* Check if process_private released. */
1318 		if (!dev->process_private)
1319 			return 0;
1320 		mlx5_tx_uar_uninit_secondary(dev);
1321 		mlx5_proc_priv_uninit(dev);
1322 		rte_eth_dev_release_port(dev);
1323 		return 0;
1324 	}
1325 	if (!priv->sh)
1326 		return 0;
1327 	DRV_LOG(DEBUG, "port %u closing device \"%s\"",
1328 		dev->data->port_id,
1329 		((priv->sh->ctx != NULL) ?
1330 		mlx5_os_get_ctx_device_name(priv->sh->ctx) : ""));
1331 	/*
1332 	 * If default mreg copy action is removed at the stop stage,
1333 	 * the search will return none and nothing will be done anymore.
1334 	 */
1335 	mlx5_flow_stop_default(dev);
1336 	mlx5_traffic_disable(dev);
1337 	/*
1338 	 * If all the flows are already flushed in the device stop stage,
1339 	 * then this will return directly without any action.
1340 	 */
1341 	mlx5_flow_list_flush(dev, &priv->flows, true);
1342 	mlx5_shared_action_flush(dev);
1343 	mlx5_flow_meter_flush(dev, NULL);
1344 	/* Prevent crashes when queues are still in use. */
1345 	dev->rx_pkt_burst = removed_rx_burst;
1346 	dev->tx_pkt_burst = removed_tx_burst;
1347 	rte_wmb();
1348 	/* Disable datapath on secondary process. */
1349 	mlx5_mp_os_req_stop_rxtx(dev);
1350 	/* Free the eCPRI flex parser resource. */
1351 	mlx5_flex_parser_ecpri_release(dev);
1352 	if (priv->rxqs != NULL) {
1353 		/* XXX race condition if mlx5_rx_burst() is still running. */
1354 		rte_delay_us_sleep(1000);
1355 		for (i = 0; (i != priv->rxqs_n); ++i)
1356 			mlx5_rxq_release(dev, i);
1357 		priv->rxqs_n = 0;
1358 		priv->rxqs = NULL;
1359 	}
1360 	if (priv->txqs != NULL) {
1361 		/* XXX race condition if mlx5_tx_burst() is still running. */
1362 		rte_delay_us_sleep(1000);
1363 		for (i = 0; (i != priv->txqs_n); ++i)
1364 			mlx5_txq_release(dev, i);
1365 		priv->txqs_n = 0;
1366 		priv->txqs = NULL;
1367 	}
1368 	mlx5_proc_priv_uninit(dev);
1369 	if (priv->q_counters) {
1370 		mlx5_devx_cmd_destroy(priv->q_counters);
1371 		priv->q_counters = NULL;
1372 	}
1373 	if (priv->drop_queue.hrxq)
1374 		mlx5_drop_action_destroy(dev);
1375 	if (priv->mreg_cp_tbl)
1376 		mlx5_hlist_destroy(priv->mreg_cp_tbl);
1377 	mlx5_mprq_free_mp(dev);
1378 	mlx5_os_free_shared_dr(priv);
1379 	if (priv->rss_conf.rss_key != NULL)
1380 		mlx5_free(priv->rss_conf.rss_key);
1381 	if (priv->reta_idx != NULL)
1382 		mlx5_free(priv->reta_idx);
1383 	if (priv->config.vf)
1384 		mlx5_os_mac_addr_flush(dev);
1385 	if (priv->nl_socket_route >= 0)
1386 		close(priv->nl_socket_route);
1387 	if (priv->nl_socket_rdma >= 0)
1388 		close(priv->nl_socket_rdma);
1389 	if (priv->vmwa_context)
1390 		mlx5_vlan_vmwa_exit(priv->vmwa_context);
1391 	ret = mlx5_hrxq_verify(dev);
1392 	if (ret)
1393 		DRV_LOG(WARNING, "port %u some hash Rx queue still remain",
1394 			dev->data->port_id);
1395 	ret = mlx5_ind_table_obj_verify(dev);
1396 	if (ret)
1397 		DRV_LOG(WARNING, "port %u some indirection table still remain",
1398 			dev->data->port_id);
1399 	ret = mlx5_rxq_obj_verify(dev);
1400 	if (ret)
1401 		DRV_LOG(WARNING, "port %u some Rx queue objects still remain",
1402 			dev->data->port_id);
1403 	ret = mlx5_rxq_verify(dev);
1404 	if (ret)
1405 		DRV_LOG(WARNING, "port %u some Rx queues still remain",
1406 			dev->data->port_id);
1407 	ret = mlx5_txq_obj_verify(dev);
1408 	if (ret)
1409 		DRV_LOG(WARNING, "port %u some Verbs Tx queue still remain",
1410 			dev->data->port_id);
1411 	ret = mlx5_txq_verify(dev);
1412 	if (ret)
1413 		DRV_LOG(WARNING, "port %u some Tx queues still remain",
1414 			dev->data->port_id);
1415 	ret = mlx5_flow_verify(dev);
1416 	if (ret)
1417 		DRV_LOG(WARNING, "port %u some flows still remain",
1418 			dev->data->port_id);
1419 	mlx5_cache_list_destroy(&priv->hrxqs);
1420 	/*
1421 	 * Free the shared context in last turn, because the cleanup
1422 	 * routines above may use some shared fields, like
1423 	 * mlx5_os_mac_addr_flush() uses ibdev_path for retrieveing
1424 	 * ifindex if Netlink fails.
1425 	 */
1426 	mlx5_free_shared_dev_ctx(priv->sh);
1427 	if (priv->domain_id != RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID) {
1428 		unsigned int c = 0;
1429 		uint16_t port_id;
1430 
1431 		MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
1432 			struct mlx5_priv *opriv =
1433 				rte_eth_devices[port_id].data->dev_private;
1434 
1435 			if (!opriv ||
1436 			    opriv->domain_id != priv->domain_id ||
1437 			    &rte_eth_devices[port_id] == dev)
1438 				continue;
1439 			++c;
1440 			break;
1441 		}
1442 		if (!c)
1443 			claim_zero(rte_eth_switch_domain_free(priv->domain_id));
1444 	}
1445 	memset(priv, 0, sizeof(*priv));
1446 	priv->domain_id = RTE_ETH_DEV_SWITCH_DOMAIN_ID_INVALID;
1447 	/*
1448 	 * Reset mac_addrs to NULL such that it is not freed as part of
1449 	 * rte_eth_dev_release_port(). mac_addrs is part of dev_private so
1450 	 * it is freed when dev_private is freed.
1451 	 */
1452 	dev->data->mac_addrs = NULL;
1453 	return 0;
1454 }
1455 
1456 const struct eth_dev_ops mlx5_dev_ops = {
1457 	.dev_configure = mlx5_dev_configure,
1458 	.dev_start = mlx5_dev_start,
1459 	.dev_stop = mlx5_dev_stop,
1460 	.dev_set_link_down = mlx5_set_link_down,
1461 	.dev_set_link_up = mlx5_set_link_up,
1462 	.dev_close = mlx5_dev_close,
1463 	.promiscuous_enable = mlx5_promiscuous_enable,
1464 	.promiscuous_disable = mlx5_promiscuous_disable,
1465 	.allmulticast_enable = mlx5_allmulticast_enable,
1466 	.allmulticast_disable = mlx5_allmulticast_disable,
1467 	.link_update = mlx5_link_update,
1468 	.stats_get = mlx5_stats_get,
1469 	.stats_reset = mlx5_stats_reset,
1470 	.xstats_get = mlx5_xstats_get,
1471 	.xstats_reset = mlx5_xstats_reset,
1472 	.xstats_get_names = mlx5_xstats_get_names,
1473 	.fw_version_get = mlx5_fw_version_get,
1474 	.dev_infos_get = mlx5_dev_infos_get,
1475 	.representor_info_get = mlx5_representor_info_get,
1476 	.read_clock = mlx5_txpp_read_clock,
1477 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1478 	.vlan_filter_set = mlx5_vlan_filter_set,
1479 	.rx_queue_setup = mlx5_rx_queue_setup,
1480 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1481 	.tx_queue_setup = mlx5_tx_queue_setup,
1482 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1483 	.rx_queue_release = mlx5_rx_queue_release,
1484 	.tx_queue_release = mlx5_tx_queue_release,
1485 	.rx_queue_start = mlx5_rx_queue_start,
1486 	.rx_queue_stop = mlx5_rx_queue_stop,
1487 	.tx_queue_start = mlx5_tx_queue_start,
1488 	.tx_queue_stop = mlx5_tx_queue_stop,
1489 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1490 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1491 	.mac_addr_remove = mlx5_mac_addr_remove,
1492 	.mac_addr_add = mlx5_mac_addr_add,
1493 	.mac_addr_set = mlx5_mac_addr_set,
1494 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1495 	.mtu_set = mlx5_dev_set_mtu,
1496 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1497 	.vlan_offload_set = mlx5_vlan_offload_set,
1498 	.reta_update = mlx5_dev_rss_reta_update,
1499 	.reta_query = mlx5_dev_rss_reta_query,
1500 	.rss_hash_update = mlx5_rss_hash_update,
1501 	.rss_hash_conf_get = mlx5_rss_hash_conf_get,
1502 	.flow_ops_get = mlx5_flow_ops_get,
1503 	.rxq_info_get = mlx5_rxq_info_get,
1504 	.txq_info_get = mlx5_txq_info_get,
1505 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1506 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1507 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1508 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1509 	.is_removed = mlx5_is_removed,
1510 	.udp_tunnel_port_add  = mlx5_udp_tunnel_port_add,
1511 	.get_module_info = mlx5_get_module_info,
1512 	.get_module_eeprom = mlx5_get_module_eeprom,
1513 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1514 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1515 	.hairpin_bind = mlx5_hairpin_bind,
1516 	.hairpin_unbind = mlx5_hairpin_unbind,
1517 	.hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1518 	.hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1519 	.hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1520 	.hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1521 };
1522 
1523 /* Available operations from secondary process. */
1524 const struct eth_dev_ops mlx5_dev_sec_ops = {
1525 	.stats_get = mlx5_stats_get,
1526 	.stats_reset = mlx5_stats_reset,
1527 	.xstats_get = mlx5_xstats_get,
1528 	.xstats_reset = mlx5_xstats_reset,
1529 	.xstats_get_names = mlx5_xstats_get_names,
1530 	.fw_version_get = mlx5_fw_version_get,
1531 	.dev_infos_get = mlx5_dev_infos_get,
1532 	.read_clock = mlx5_txpp_read_clock,
1533 	.rx_queue_start = mlx5_rx_queue_start,
1534 	.rx_queue_stop = mlx5_rx_queue_stop,
1535 	.tx_queue_start = mlx5_tx_queue_start,
1536 	.tx_queue_stop = mlx5_tx_queue_stop,
1537 	.rxq_info_get = mlx5_rxq_info_get,
1538 	.txq_info_get = mlx5_txq_info_get,
1539 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1540 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1541 	.get_module_info = mlx5_get_module_info,
1542 	.get_module_eeprom = mlx5_get_module_eeprom,
1543 };
1544 
1545 /* Available operations in flow isolated mode. */
1546 const struct eth_dev_ops mlx5_dev_ops_isolate = {
1547 	.dev_configure = mlx5_dev_configure,
1548 	.dev_start = mlx5_dev_start,
1549 	.dev_stop = mlx5_dev_stop,
1550 	.dev_set_link_down = mlx5_set_link_down,
1551 	.dev_set_link_up = mlx5_set_link_up,
1552 	.dev_close = mlx5_dev_close,
1553 	.promiscuous_enable = mlx5_promiscuous_enable,
1554 	.promiscuous_disable = mlx5_promiscuous_disable,
1555 	.allmulticast_enable = mlx5_allmulticast_enable,
1556 	.allmulticast_disable = mlx5_allmulticast_disable,
1557 	.link_update = mlx5_link_update,
1558 	.stats_get = mlx5_stats_get,
1559 	.stats_reset = mlx5_stats_reset,
1560 	.xstats_get = mlx5_xstats_get,
1561 	.xstats_reset = mlx5_xstats_reset,
1562 	.xstats_get_names = mlx5_xstats_get_names,
1563 	.fw_version_get = mlx5_fw_version_get,
1564 	.dev_infos_get = mlx5_dev_infos_get,
1565 	.read_clock = mlx5_txpp_read_clock,
1566 	.dev_supported_ptypes_get = mlx5_dev_supported_ptypes_get,
1567 	.vlan_filter_set = mlx5_vlan_filter_set,
1568 	.rx_queue_setup = mlx5_rx_queue_setup,
1569 	.rx_hairpin_queue_setup = mlx5_rx_hairpin_queue_setup,
1570 	.tx_queue_setup = mlx5_tx_queue_setup,
1571 	.tx_hairpin_queue_setup = mlx5_tx_hairpin_queue_setup,
1572 	.rx_queue_release = mlx5_rx_queue_release,
1573 	.tx_queue_release = mlx5_tx_queue_release,
1574 	.rx_queue_start = mlx5_rx_queue_start,
1575 	.rx_queue_stop = mlx5_rx_queue_stop,
1576 	.tx_queue_start = mlx5_tx_queue_start,
1577 	.tx_queue_stop = mlx5_tx_queue_stop,
1578 	.flow_ctrl_get = mlx5_dev_get_flow_ctrl,
1579 	.flow_ctrl_set = mlx5_dev_set_flow_ctrl,
1580 	.mac_addr_remove = mlx5_mac_addr_remove,
1581 	.mac_addr_add = mlx5_mac_addr_add,
1582 	.mac_addr_set = mlx5_mac_addr_set,
1583 	.set_mc_addr_list = mlx5_set_mc_addr_list,
1584 	.mtu_set = mlx5_dev_set_mtu,
1585 	.vlan_strip_queue_set = mlx5_vlan_strip_queue_set,
1586 	.vlan_offload_set = mlx5_vlan_offload_set,
1587 	.flow_ops_get = mlx5_flow_ops_get,
1588 	.rxq_info_get = mlx5_rxq_info_get,
1589 	.txq_info_get = mlx5_txq_info_get,
1590 	.rx_burst_mode_get = mlx5_rx_burst_mode_get,
1591 	.tx_burst_mode_get = mlx5_tx_burst_mode_get,
1592 	.rx_queue_intr_enable = mlx5_rx_intr_enable,
1593 	.rx_queue_intr_disable = mlx5_rx_intr_disable,
1594 	.is_removed = mlx5_is_removed,
1595 	.get_module_info = mlx5_get_module_info,
1596 	.get_module_eeprom = mlx5_get_module_eeprom,
1597 	.hairpin_cap_get = mlx5_hairpin_cap_get,
1598 	.mtr_ops_get = mlx5_flow_meter_ops_get,
1599 	.hairpin_bind = mlx5_hairpin_bind,
1600 	.hairpin_unbind = mlx5_hairpin_unbind,
1601 	.hairpin_get_peer_ports = mlx5_hairpin_get_peer_ports,
1602 	.hairpin_queue_peer_update = mlx5_hairpin_queue_peer_update,
1603 	.hairpin_queue_peer_bind = mlx5_hairpin_queue_peer_bind,
1604 	.hairpin_queue_peer_unbind = mlx5_hairpin_queue_peer_unbind,
1605 };
1606 
1607 /**
1608  * Verify and store value for device argument.
1609  *
1610  * @param[in] key
1611  *   Key argument to verify.
1612  * @param[in] val
1613  *   Value associated with key.
1614  * @param opaque
1615  *   User data.
1616  *
1617  * @return
1618  *   0 on success, a negative errno value otherwise and rte_errno is set.
1619  */
1620 static int
1621 mlx5_args_check(const char *key, const char *val, void *opaque)
1622 {
1623 	struct mlx5_dev_config *config = opaque;
1624 	unsigned long mod;
1625 	signed long tmp;
1626 
1627 	/* No-op, port representors are processed in mlx5_dev_spawn(). */
1628 	if (!strcmp(MLX5_REPRESENTOR, key))
1629 		return 0;
1630 	errno = 0;
1631 	tmp = strtol(val, NULL, 0);
1632 	if (errno) {
1633 		rte_errno = errno;
1634 		DRV_LOG(WARNING, "%s: \"%s\" is not a valid integer", key, val);
1635 		return -rte_errno;
1636 	}
1637 	if (tmp < 0 && strcmp(MLX5_TX_PP, key) && strcmp(MLX5_TX_SKEW, key)) {
1638 		/* Negative values are acceptable for some keys only. */
1639 		rte_errno = EINVAL;
1640 		DRV_LOG(WARNING, "%s: invalid negative value \"%s\"", key, val);
1641 		return -rte_errno;
1642 	}
1643 	mod = tmp >= 0 ? tmp : -tmp;
1644 	if (strcmp(MLX5_RXQ_CQE_COMP_EN, key) == 0) {
1645 		if (tmp > MLX5_CQE_RESP_FORMAT_L34H_STRIDX) {
1646 			DRV_LOG(ERR, "invalid CQE compression "
1647 				     "format parameter");
1648 			rte_errno = EINVAL;
1649 			return -rte_errno;
1650 		}
1651 		config->cqe_comp = !!tmp;
1652 		config->cqe_comp_fmt = tmp;
1653 	} else if (strcmp(MLX5_RXQ_PKT_PAD_EN, key) == 0) {
1654 		config->hw_padding = !!tmp;
1655 	} else if (strcmp(MLX5_RX_MPRQ_EN, key) == 0) {
1656 		config->mprq.enabled = !!tmp;
1657 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_NUM, key) == 0) {
1658 		config->mprq.stride_num_n = tmp;
1659 	} else if (strcmp(MLX5_RX_MPRQ_LOG_STRIDE_SIZE, key) == 0) {
1660 		config->mprq.stride_size_n = tmp;
1661 	} else if (strcmp(MLX5_RX_MPRQ_MAX_MEMCPY_LEN, key) == 0) {
1662 		config->mprq.max_memcpy_len = tmp;
1663 	} else if (strcmp(MLX5_RXQS_MIN_MPRQ, key) == 0) {
1664 		config->mprq.min_rxqs_num = tmp;
1665 	} else if (strcmp(MLX5_TXQ_INLINE, key) == 0) {
1666 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1667 				 " converted to txq_inline_max", key);
1668 		config->txq_inline_max = tmp;
1669 	} else if (strcmp(MLX5_TXQ_INLINE_MAX, key) == 0) {
1670 		config->txq_inline_max = tmp;
1671 	} else if (strcmp(MLX5_TXQ_INLINE_MIN, key) == 0) {
1672 		config->txq_inline_min = tmp;
1673 	} else if (strcmp(MLX5_TXQ_INLINE_MPW, key) == 0) {
1674 		config->txq_inline_mpw = tmp;
1675 	} else if (strcmp(MLX5_TXQS_MIN_INLINE, key) == 0) {
1676 		config->txqs_inline = tmp;
1677 	} else if (strcmp(MLX5_TXQS_MAX_VEC, key) == 0) {
1678 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1679 	} else if (strcmp(MLX5_TXQ_MPW_EN, key) == 0) {
1680 		config->mps = !!tmp;
1681 	} else if (strcmp(MLX5_TX_DB_NC, key) == 0) {
1682 		if (tmp != MLX5_TXDB_CACHED &&
1683 		    tmp != MLX5_TXDB_NCACHED &&
1684 		    tmp != MLX5_TXDB_HEURISTIC) {
1685 			DRV_LOG(ERR, "invalid Tx doorbell "
1686 				     "mapping parameter");
1687 			rte_errno = EINVAL;
1688 			return -rte_errno;
1689 		}
1690 		config->dbnc = tmp;
1691 	} else if (strcmp(MLX5_TXQ_MPW_HDR_DSEG_EN, key) == 0) {
1692 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1693 	} else if (strcmp(MLX5_TXQ_MAX_INLINE_LEN, key) == 0) {
1694 		DRV_LOG(WARNING, "%s: deprecated parameter,"
1695 				 " converted to txq_inline_mpw", key);
1696 		config->txq_inline_mpw = tmp;
1697 	} else if (strcmp(MLX5_TX_VEC_EN, key) == 0) {
1698 		DRV_LOG(WARNING, "%s: deprecated parameter, ignored", key);
1699 	} else if (strcmp(MLX5_TX_PP, key) == 0) {
1700 		if (!mod) {
1701 			DRV_LOG(ERR, "Zero Tx packet pacing parameter");
1702 			rte_errno = EINVAL;
1703 			return -rte_errno;
1704 		}
1705 		config->tx_pp = tmp;
1706 	} else if (strcmp(MLX5_TX_SKEW, key) == 0) {
1707 		config->tx_skew = tmp;
1708 	} else if (strcmp(MLX5_RX_VEC_EN, key) == 0) {
1709 		config->rx_vec_en = !!tmp;
1710 	} else if (strcmp(MLX5_L3_VXLAN_EN, key) == 0) {
1711 		config->l3_vxlan_en = !!tmp;
1712 	} else if (strcmp(MLX5_VF_NL_EN, key) == 0) {
1713 		config->vf_nl_en = !!tmp;
1714 	} else if (strcmp(MLX5_DV_ESW_EN, key) == 0) {
1715 		config->dv_esw_en = !!tmp;
1716 	} else if (strcmp(MLX5_DV_FLOW_EN, key) == 0) {
1717 		config->dv_flow_en = !!tmp;
1718 	} else if (strcmp(MLX5_DV_XMETA_EN, key) == 0) {
1719 		if (tmp != MLX5_XMETA_MODE_LEGACY &&
1720 		    tmp != MLX5_XMETA_MODE_META16 &&
1721 		    tmp != MLX5_XMETA_MODE_META32 &&
1722 		    tmp != MLX5_XMETA_MODE_MISS_INFO) {
1723 			DRV_LOG(ERR, "invalid extensive "
1724 				     "metadata parameter");
1725 			rte_errno = EINVAL;
1726 			return -rte_errno;
1727 		}
1728 		if (tmp != MLX5_XMETA_MODE_MISS_INFO)
1729 			config->dv_xmeta_en = tmp;
1730 		else
1731 			config->dv_miss_info = 1;
1732 	} else if (strcmp(MLX5_LACP_BY_USER, key) == 0) {
1733 		config->lacp_by_user = !!tmp;
1734 	} else if (strcmp(MLX5_MR_EXT_MEMSEG_EN, key) == 0) {
1735 		config->mr_ext_memseg_en = !!tmp;
1736 	} else if (strcmp(MLX5_MAX_DUMP_FILES_NUM, key) == 0) {
1737 		config->max_dump_files_num = tmp;
1738 	} else if (strcmp(MLX5_LRO_TIMEOUT_USEC, key) == 0) {
1739 		config->lro.timeout = tmp;
1740 	} else if (strcmp(MLX5_CLASS_ARG_NAME, key) == 0) {
1741 		DRV_LOG(DEBUG, "class argument is %s.", val);
1742 	} else if (strcmp(MLX5_HP_BUF_SIZE, key) == 0) {
1743 		config->log_hp_size = tmp;
1744 	} else if (strcmp(MLX5_RECLAIM_MEM, key) == 0) {
1745 		if (tmp != MLX5_RCM_NONE &&
1746 		    tmp != MLX5_RCM_LIGHT &&
1747 		    tmp != MLX5_RCM_AGGR) {
1748 			DRV_LOG(ERR, "Unrecognize %s: \"%s\"", key, val);
1749 			rte_errno = EINVAL;
1750 			return -rte_errno;
1751 		}
1752 		config->reclaim_mode = tmp;
1753 	} else if (strcmp(MLX5_SYS_MEM_EN, key) == 0) {
1754 		config->sys_mem_en = !!tmp;
1755 	} else if (strcmp(MLX5_DECAP_EN, key) == 0) {
1756 		config->decap_en = !!tmp;
1757 	} else {
1758 		DRV_LOG(WARNING, "%s: unknown parameter", key);
1759 		rte_errno = EINVAL;
1760 		return -rte_errno;
1761 	}
1762 	return 0;
1763 }
1764 
1765 /**
1766  * Parse device parameters.
1767  *
1768  * @param config
1769  *   Pointer to device configuration structure.
1770  * @param devargs
1771  *   Device arguments structure.
1772  *
1773  * @return
1774  *   0 on success, a negative errno value otherwise and rte_errno is set.
1775  */
1776 int
1777 mlx5_args(struct mlx5_dev_config *config, struct rte_devargs *devargs)
1778 {
1779 	const char **params = (const char *[]){
1780 		MLX5_RXQ_CQE_COMP_EN,
1781 		MLX5_RXQ_PKT_PAD_EN,
1782 		MLX5_RX_MPRQ_EN,
1783 		MLX5_RX_MPRQ_LOG_STRIDE_NUM,
1784 		MLX5_RX_MPRQ_LOG_STRIDE_SIZE,
1785 		MLX5_RX_MPRQ_MAX_MEMCPY_LEN,
1786 		MLX5_RXQS_MIN_MPRQ,
1787 		MLX5_TXQ_INLINE,
1788 		MLX5_TXQ_INLINE_MIN,
1789 		MLX5_TXQ_INLINE_MAX,
1790 		MLX5_TXQ_INLINE_MPW,
1791 		MLX5_TXQS_MIN_INLINE,
1792 		MLX5_TXQS_MAX_VEC,
1793 		MLX5_TXQ_MPW_EN,
1794 		MLX5_TXQ_MPW_HDR_DSEG_EN,
1795 		MLX5_TXQ_MAX_INLINE_LEN,
1796 		MLX5_TX_DB_NC,
1797 		MLX5_TX_PP,
1798 		MLX5_TX_SKEW,
1799 		MLX5_TX_VEC_EN,
1800 		MLX5_RX_VEC_EN,
1801 		MLX5_L3_VXLAN_EN,
1802 		MLX5_VF_NL_EN,
1803 		MLX5_DV_ESW_EN,
1804 		MLX5_DV_FLOW_EN,
1805 		MLX5_DV_XMETA_EN,
1806 		MLX5_LACP_BY_USER,
1807 		MLX5_MR_EXT_MEMSEG_EN,
1808 		MLX5_REPRESENTOR,
1809 		MLX5_MAX_DUMP_FILES_NUM,
1810 		MLX5_LRO_TIMEOUT_USEC,
1811 		MLX5_CLASS_ARG_NAME,
1812 		MLX5_HP_BUF_SIZE,
1813 		MLX5_RECLAIM_MEM,
1814 		MLX5_SYS_MEM_EN,
1815 		MLX5_DECAP_EN,
1816 		NULL,
1817 	};
1818 	struct rte_kvargs *kvlist;
1819 	int ret = 0;
1820 	int i;
1821 
1822 	if (devargs == NULL)
1823 		return 0;
1824 	/* Following UGLY cast is done to pass checkpatch. */
1825 	kvlist = rte_kvargs_parse(devargs->args, params);
1826 	if (kvlist == NULL) {
1827 		rte_errno = EINVAL;
1828 		return -rte_errno;
1829 	}
1830 	/* Process parameters. */
1831 	for (i = 0; (params[i] != NULL); ++i) {
1832 		if (rte_kvargs_count(kvlist, params[i])) {
1833 			ret = rte_kvargs_process(kvlist, params[i],
1834 						 mlx5_args_check, config);
1835 			if (ret) {
1836 				rte_errno = EINVAL;
1837 				rte_kvargs_free(kvlist);
1838 				return -rte_errno;
1839 			}
1840 		}
1841 	}
1842 	rte_kvargs_free(kvlist);
1843 	return 0;
1844 }
1845 
1846 /**
1847  * Configures the minimal amount of data to inline into WQE
1848  * while sending packets.
1849  *
1850  * - the txq_inline_min has the maximal priority, if this
1851  *   key is specified in devargs
1852  * - if DevX is enabled the inline mode is queried from the
1853  *   device (HCA attributes and NIC vport context if needed).
1854  * - otherwise L2 mode (18 bytes) is assumed for ConnectX-4/4 Lx
1855  *   and none (0 bytes) for other NICs
1856  *
1857  * @param spawn
1858  *   Verbs device parameters (name, port, switch_info) to spawn.
1859  * @param config
1860  *   Device configuration parameters.
1861  */
1862 void
1863 mlx5_set_min_inline(struct mlx5_dev_spawn_data *spawn,
1864 		    struct mlx5_dev_config *config)
1865 {
1866 	if (config->txq_inline_min != MLX5_ARG_UNSET) {
1867 		/* Application defines size of inlined data explicitly. */
1868 		switch (spawn->pci_dev->id.device_id) {
1869 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1870 		case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1871 			if (config->txq_inline_min <
1872 				       (int)MLX5_INLINE_HSIZE_L2) {
1873 				DRV_LOG(DEBUG,
1874 					"txq_inline_mix aligned to minimal"
1875 					" ConnectX-4 required value %d",
1876 					(int)MLX5_INLINE_HSIZE_L2);
1877 				config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1878 			}
1879 			break;
1880 		}
1881 		goto exit;
1882 	}
1883 	if (config->hca_attr.eth_net_offloads) {
1884 		/* We have DevX enabled, inline mode queried successfully. */
1885 		switch (config->hca_attr.wqe_inline_mode) {
1886 		case MLX5_CAP_INLINE_MODE_L2:
1887 			/* outer L2 header must be inlined. */
1888 			config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1889 			goto exit;
1890 		case MLX5_CAP_INLINE_MODE_NOT_REQUIRED:
1891 			/* No inline data are required by NIC. */
1892 			config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1893 			config->hw_vlan_insert =
1894 				config->hca_attr.wqe_vlan_insert;
1895 			DRV_LOG(DEBUG, "Tx VLAN insertion is supported");
1896 			goto exit;
1897 		case MLX5_CAP_INLINE_MODE_VPORT_CONTEXT:
1898 			/* inline mode is defined by NIC vport context. */
1899 			if (!config->hca_attr.eth_virt)
1900 				break;
1901 			switch (config->hca_attr.vport_inline_mode) {
1902 			case MLX5_INLINE_MODE_NONE:
1903 				config->txq_inline_min =
1904 					MLX5_INLINE_HSIZE_NONE;
1905 				goto exit;
1906 			case MLX5_INLINE_MODE_L2:
1907 				config->txq_inline_min =
1908 					MLX5_INLINE_HSIZE_L2;
1909 				goto exit;
1910 			case MLX5_INLINE_MODE_IP:
1911 				config->txq_inline_min =
1912 					MLX5_INLINE_HSIZE_L3;
1913 				goto exit;
1914 			case MLX5_INLINE_MODE_TCP_UDP:
1915 				config->txq_inline_min =
1916 					MLX5_INLINE_HSIZE_L4;
1917 				goto exit;
1918 			case MLX5_INLINE_MODE_INNER_L2:
1919 				config->txq_inline_min =
1920 					MLX5_INLINE_HSIZE_INNER_L2;
1921 				goto exit;
1922 			case MLX5_INLINE_MODE_INNER_IP:
1923 				config->txq_inline_min =
1924 					MLX5_INLINE_HSIZE_INNER_L3;
1925 				goto exit;
1926 			case MLX5_INLINE_MODE_INNER_TCP_UDP:
1927 				config->txq_inline_min =
1928 					MLX5_INLINE_HSIZE_INNER_L4;
1929 				goto exit;
1930 			}
1931 		}
1932 	}
1933 	/*
1934 	 * We get here if we are unable to deduce
1935 	 * inline data size with DevX. Try PCI ID
1936 	 * to determine old NICs.
1937 	 */
1938 	switch (spawn->pci_dev->id.device_id) {
1939 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4:
1940 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4VF:
1941 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LX:
1942 	case PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF:
1943 		config->txq_inline_min = MLX5_INLINE_HSIZE_L2;
1944 		config->hw_vlan_insert = 0;
1945 		break;
1946 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5:
1947 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5VF:
1948 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EX:
1949 	case PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF:
1950 		/*
1951 		 * These NICs support VLAN insertion from WQE and
1952 		 * report the wqe_vlan_insert flag. But there is the bug
1953 		 * and PFC control may be broken, so disable feature.
1954 		 */
1955 		config->hw_vlan_insert = 0;
1956 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1957 		break;
1958 	default:
1959 		config->txq_inline_min = MLX5_INLINE_HSIZE_NONE;
1960 		break;
1961 	}
1962 exit:
1963 	DRV_LOG(DEBUG, "min tx inline configured: %d", config->txq_inline_min);
1964 }
1965 
1966 /**
1967  * Configures the metadata mask fields in the shared context.
1968  *
1969  * @param [in] dev
1970  *   Pointer to Ethernet device.
1971  */
1972 void
1973 mlx5_set_metadata_mask(struct rte_eth_dev *dev)
1974 {
1975 	struct mlx5_priv *priv = dev->data->dev_private;
1976 	struct mlx5_dev_ctx_shared *sh = priv->sh;
1977 	uint32_t meta, mark, reg_c0;
1978 
1979 	reg_c0 = ~priv->vport_meta_mask;
1980 	switch (priv->config.dv_xmeta_en) {
1981 	case MLX5_XMETA_MODE_LEGACY:
1982 		meta = UINT32_MAX;
1983 		mark = MLX5_FLOW_MARK_MASK;
1984 		break;
1985 	case MLX5_XMETA_MODE_META16:
1986 		meta = reg_c0 >> rte_bsf32(reg_c0);
1987 		mark = MLX5_FLOW_MARK_MASK;
1988 		break;
1989 	case MLX5_XMETA_MODE_META32:
1990 		meta = UINT32_MAX;
1991 		mark = (reg_c0 >> rte_bsf32(reg_c0)) & MLX5_FLOW_MARK_MASK;
1992 		break;
1993 	default:
1994 		meta = 0;
1995 		mark = 0;
1996 		MLX5_ASSERT(false);
1997 		break;
1998 	}
1999 	if (sh->dv_mark_mask && sh->dv_mark_mask != mark)
2000 		DRV_LOG(WARNING, "metadata MARK mask mismatche %08X:%08X",
2001 				 sh->dv_mark_mask, mark);
2002 	else
2003 		sh->dv_mark_mask = mark;
2004 	if (sh->dv_meta_mask && sh->dv_meta_mask != meta)
2005 		DRV_LOG(WARNING, "metadata META mask mismatche %08X:%08X",
2006 				 sh->dv_meta_mask, meta);
2007 	else
2008 		sh->dv_meta_mask = meta;
2009 	if (sh->dv_regc0_mask && sh->dv_regc0_mask != reg_c0)
2010 		DRV_LOG(WARNING, "metadata reg_c0 mask mismatche %08X:%08X",
2011 				 sh->dv_meta_mask, reg_c0);
2012 	else
2013 		sh->dv_regc0_mask = reg_c0;
2014 	DRV_LOG(DEBUG, "metadata mode %u", priv->config.dv_xmeta_en);
2015 	DRV_LOG(DEBUG, "metadata MARK mask %08X", sh->dv_mark_mask);
2016 	DRV_LOG(DEBUG, "metadata META mask %08X", sh->dv_meta_mask);
2017 	DRV_LOG(DEBUG, "metadata reg_c0 mask %08X", sh->dv_regc0_mask);
2018 }
2019 
2020 int
2021 rte_pmd_mlx5_get_dyn_flag_names(char *names[], unsigned int n)
2022 {
2023 	static const char *const dynf_names[] = {
2024 		RTE_PMD_MLX5_FINE_GRANULARITY_INLINE,
2025 		RTE_MBUF_DYNFLAG_METADATA_NAME,
2026 		RTE_MBUF_DYNFLAG_TX_TIMESTAMP_NAME
2027 	};
2028 	unsigned int i;
2029 
2030 	if (n < RTE_DIM(dynf_names))
2031 		return -ENOMEM;
2032 	for (i = 0; i < RTE_DIM(dynf_names); i++) {
2033 		if (names[i] == NULL)
2034 			return -EINVAL;
2035 		strcpy(names[i], dynf_names[i]);
2036 	}
2037 	return RTE_DIM(dynf_names);
2038 }
2039 
2040 /**
2041  * Comparison callback to sort device data.
2042  *
2043  * This is meant to be used with qsort().
2044  *
2045  * @param a[in]
2046  *   Pointer to pointer to first data object.
2047  * @param b[in]
2048  *   Pointer to pointer to second data object.
2049  *
2050  * @return
2051  *   0 if both objects are equal, less than 0 if the first argument is less
2052  *   than the second, greater than 0 otherwise.
2053  */
2054 int
2055 mlx5_dev_check_sibling_config(struct mlx5_priv *priv,
2056 			      struct mlx5_dev_config *config)
2057 {
2058 	struct mlx5_dev_ctx_shared *sh = priv->sh;
2059 	struct mlx5_dev_config *sh_conf = NULL;
2060 	uint16_t port_id;
2061 
2062 	MLX5_ASSERT(sh);
2063 	/* Nothing to compare for the single/first device. */
2064 	if (sh->refcnt == 1)
2065 		return 0;
2066 	/* Find the device with shared context. */
2067 	MLX5_ETH_FOREACH_DEV(port_id, priv->pci_dev) {
2068 		struct mlx5_priv *opriv =
2069 			rte_eth_devices[port_id].data->dev_private;
2070 
2071 		if (opriv && opriv != priv && opriv->sh == sh) {
2072 			sh_conf = &opriv->config;
2073 			break;
2074 		}
2075 	}
2076 	if (!sh_conf)
2077 		return 0;
2078 	if (sh_conf->dv_flow_en ^ config->dv_flow_en) {
2079 		DRV_LOG(ERR, "\"dv_flow_en\" configuration mismatch"
2080 			     " for shared %s context", sh->ibdev_name);
2081 		rte_errno = EINVAL;
2082 		return rte_errno;
2083 	}
2084 	if (sh_conf->dv_xmeta_en ^ config->dv_xmeta_en) {
2085 		DRV_LOG(ERR, "\"dv_xmeta_en\" configuration mismatch"
2086 			     " for shared %s context", sh->ibdev_name);
2087 		rte_errno = EINVAL;
2088 		return rte_errno;
2089 	}
2090 	return 0;
2091 }
2092 
2093 /**
2094  * Look for the ethernet device belonging to mlx5 driver.
2095  *
2096  * @param[in] port_id
2097  *   port_id to start looking for device.
2098  * @param[in] pci_dev
2099  *   Pointer to the hint PCI device. When device is being probed
2100  *   the its siblings (master and preceding representors might
2101  *   not have assigned driver yet (because the mlx5_os_pci_probe()
2102  *   is not completed yet, for this case match on hint PCI
2103  *   device may be used to detect sibling device.
2104  *
2105  * @return
2106  *   port_id of found device, RTE_MAX_ETHPORT if not found.
2107  */
2108 uint16_t
2109 mlx5_eth_find_next(uint16_t port_id, struct rte_pci_device *pci_dev)
2110 {
2111 	while (port_id < RTE_MAX_ETHPORTS) {
2112 		struct rte_eth_dev *dev = &rte_eth_devices[port_id];
2113 
2114 		if (dev->state != RTE_ETH_DEV_UNUSED &&
2115 		    dev->device &&
2116 		    (dev->device == &pci_dev->device ||
2117 		     (dev->device->driver &&
2118 		     dev->device->driver->name &&
2119 		     !strcmp(dev->device->driver->name, MLX5_DRIVER_NAME))))
2120 			break;
2121 		port_id++;
2122 	}
2123 	if (port_id >= RTE_MAX_ETHPORTS)
2124 		return RTE_MAX_ETHPORTS;
2125 	return port_id;
2126 }
2127 
2128 /**
2129  * DPDK callback to remove a PCI device.
2130  *
2131  * This function removes all Ethernet devices belong to a given PCI device.
2132  *
2133  * @param[in] pci_dev
2134  *   Pointer to the PCI device.
2135  *
2136  * @return
2137  *   0 on success, the function cannot fail.
2138  */
2139 static int
2140 mlx5_pci_remove(struct rte_pci_device *pci_dev)
2141 {
2142 	uint16_t port_id;
2143 	int ret = 0;
2144 
2145 	RTE_ETH_FOREACH_DEV_OF(port_id, &pci_dev->device) {
2146 		/*
2147 		 * mlx5_dev_close() is not registered to secondary process,
2148 		 * call the close function explicitly for secondary process.
2149 		 */
2150 		if (rte_eal_process_type() == RTE_PROC_SECONDARY)
2151 			ret |= mlx5_dev_close(&rte_eth_devices[port_id]);
2152 		else
2153 			ret |= rte_eth_dev_close(port_id);
2154 	}
2155 	return ret == 0 ? 0 : -EIO;
2156 }
2157 
2158 static const struct rte_pci_id mlx5_pci_id_map[] = {
2159 	{
2160 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2161 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4)
2162 	},
2163 	{
2164 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2165 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4VF)
2166 	},
2167 	{
2168 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2169 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LX)
2170 	},
2171 	{
2172 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2173 			       PCI_DEVICE_ID_MELLANOX_CONNECTX4LXVF)
2174 	},
2175 	{
2176 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2177 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5)
2178 	},
2179 	{
2180 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2181 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5VF)
2182 	},
2183 	{
2184 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2185 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EX)
2186 	},
2187 	{
2188 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2189 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5EXVF)
2190 	},
2191 	{
2192 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2193 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
2194 	},
2195 	{
2196 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2197 			       PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
2198 	},
2199 	{
2200 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2201 				PCI_DEVICE_ID_MELLANOX_CONNECTX6)
2202 	},
2203 	{
2204 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2205 				PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
2206 	},
2207 	{
2208 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2209 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
2210 	},
2211 	{
2212 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2213 				PCI_DEVICE_ID_MELLANOX_CONNECTXVF)
2214 	},
2215 	{
2216 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2217 				PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
2218 	},
2219 	{
2220 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2221 				PCI_DEVICE_ID_MELLANOX_CONNECTX6LX)
2222 	},
2223 	{
2224 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2225 				PCI_DEVICE_ID_MELLANOX_CONNECTX7)
2226 	},
2227 	{
2228 		RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
2229 				PCI_DEVICE_ID_MELLANOX_CONNECTX7BF)
2230 	},
2231 	{
2232 		.vendor_id = 0
2233 	}
2234 };
2235 
2236 static struct mlx5_pci_driver mlx5_driver = {
2237 	.driver_class = MLX5_CLASS_NET,
2238 	.pci_driver = {
2239 		.driver = {
2240 			.name = MLX5_DRIVER_NAME,
2241 		},
2242 		.id_table = mlx5_pci_id_map,
2243 		.probe = mlx5_os_pci_probe,
2244 		.remove = mlx5_pci_remove,
2245 		.dma_map = mlx5_dma_map,
2246 		.dma_unmap = mlx5_dma_unmap,
2247 		.drv_flags = PCI_DRV_FLAGS,
2248 	},
2249 };
2250 
2251 /* Initialize driver log type. */
2252 RTE_LOG_REGISTER(mlx5_logtype, pmd.net.mlx5, NOTICE)
2253 
2254 /**
2255  * Driver initialization routine.
2256  */
2257 RTE_INIT(rte_mlx5_pmd_init)
2258 {
2259 	pthread_mutex_init(&mlx5_dev_ctx_list_mutex, NULL);
2260 	mlx5_common_init();
2261 	/* Build the static tables for Verbs conversion. */
2262 	mlx5_set_ptype_table();
2263 	mlx5_set_cksum_table();
2264 	mlx5_set_swp_types_table();
2265 	if (mlx5_glue)
2266 		mlx5_pci_driver_register(&mlx5_driver);
2267 }
2268 
2269 RTE_PMD_EXPORT_NAME(net_mlx5, __COUNTER__);
2270 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5, mlx5_pci_id_map);
2271 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5, "* ib_uverbs & mlx5_core & mlx5_ib");
2272